Encoder-only transformer with simulation-to-experiment transfer for radionuclide identification in NaI(Tl) spectroscopy

Radionuclide identification is central to nuclear safety, security, and environmental monitoring, yet NaI(Tl) detectors remain limited by low resolution, peak overlap, background variability, and simulation-experiment discrepancies. Existing neural-network models typically achieve high accuracy on simulated spectra but generalize poorly to experimental measurements due to domain shift. This study proposes an encoder-only transformer for radionuclide identification using NaI(Tl) spectra. The model is trained and fine-tuned on experimentally measured data. By leveraging self-attention, it captures global spectral dependencies and consistently outperforms conventional neural networks, such as Convolutional Neural Network, Residual Network, and Multilayer Perceptron in realistic measurement scenarios. This approach offers a generalizable framework for reliable and efficient radionuclide identification, supporting broader applications in nuclear safety, security, and environmental monitoring. A calibrated Geant4 model provides simulated spectra for seven radionuclides ( 241 Am, 226 Ra, 57 Co, 60 Co, 137 Cs, 152 Eu, 133 Ba), augmented by synthetic mixture generation. Experimental spectra were acquired of four radionuclides ( 60 Co, 137 Cs, 152 Eu, 133 Ba) including all 15 source combinations from single- to four-nuclide mixtures. The model is pretrained on simulated spectra and fine-tuned with a limited amount of experimental data, while all remaining spectra are reserved for testing. It consistently outperforms traditional peak analysis method and neural networks (Convolutional Neural Network, Residual Network, and Multilayer Perceptron) and achieves about 95% average accuracy on simulated spectra and 93% on experimental spectra, surpassing the strongest baseline by over 3 percentage points. Fine-tuning yields consistent improvements, especially for complex triple-nuclide mixtures, confirming the robustness and data efficiency of the model under realistic NaI(Tl) measurement conditions. Ablation results confirm that both the encoder-only architecture and the simulation-to-experiment training strategy contribute critically to performance gains, while efficiency analysis shows that the model achieves millisecond-level inference suitable for practical deployment.